Neural network river forecasting through baseflow separation and binary-coded swarm optimization Riccardo Taormina a , Kwok-Wing Chau a,⇑ , Bellie Sivakumar b,c a Department of Civil and Environmental Engineering, Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong b School of Civil and Environmental Engineering, The University of New South Wales, Sydney, NSW 2052, Australia c Department of Land, Air and Water Resources, University of California, Davis, CA 95616, USA article info Article history: Received 8 June 2015 Received in revised form 7 August 2015 Accepted 8 August 2015 Available online 15 August 2015 This manuscript was handled by Geoff Syme, Editor-in-Chief Keywords: Rainfall–runoff Baseflow separation Extreme Learning Machine Multi-objective optimization Particle swarm optimization Modular neural network summary The inclusion of expert knowledge in data-driven streamflow modeling is expected to yield more accurate estimates of river quantities. Modular models (MMs) designed to work on different parts of the hydrograph are preferred ways to implement such approach. Previous studies have suggested that better predictions of total streamflow could be obtained via modular Artificial Neural Networks (ANNs) trained to perform an implicit baseflow separation. These MMs fit separately the baseflow and excess flow components as produced by a digital filter, and reconstruct the total flow by adding these two signals at the output. The optimization of the filter parameters and ANN architectures is carried out through global search techniques. Despite the favorable premises, the real effectiveness of such MMs has been tested only on a few case studies, and the quality of the baseflow separation they perform has never been thoroughly assessed. In this work, we compare the performance of MM against global models (GMs) for nine different gaging stations in the northern United States. Binary-coded swarm opti- mization is employed for the identification of filter parameters and model structure, while Extreme Learning Machines, instead of ANN, are used to drastically reduce the large computational times required to perform the experiments. The results show that there is no evidence that MM outperform global GM for predicting the total flow. In addition, the baseflow produced by the MM largely underestimates the actual baseflow component expected for most of the considered gages. This occurs because the values of the filter parameters maximizing overall accuracy do not reflect the geological characteristics of the river basins. The results indeed show that setting the filter parameters according to expert knowledge results in accurate baseflow separation but lower accuracy of total flow predictions, suggesting that these two objectives are intrinsically conflicting rather than compatible. Ó 2015 Elsevier B.V. All rights reserved. 1. Introduction In the past few decades, data-driven modeling tools have been widely employed for research in hydrological modeling; see Sivakumar and Berndtsson (2010) for a comprehensive account of some of the major data-based approaches in hydrology. In par- ticular, Artificial Neural Networks (ANNs) have been extensively used due to their ability to cope with the non-linear, non- stationary and non-Gaussian processes typically found in hydrol- ogy (ASCE Task Committee, 2000; Maier and Dandy, 2000; Maier et al., 2010), as well as for their flexibility and relative ease in development. ANN applications in hydrology range from water quality modeling (Wu et al., 2014) and rainfall forecasting (Wu et al., 2010), to groundwater simulation (Taormina et al., 2012) and reservoir operations (Jothiprakash and Magar, 2012). However, these techniques have been primarily employed for rainfall–runoff modeling and streamflow prediction (Abrahart et al., 2012; Corzo and Solomatine, 2007a, 2007b; de Vos and Rientjes, 2008; Jain and Srinivasulu, 2006; See and Openshaw, 1999; Srinivasulu and Jain, 2009; Taormina and Chau, 2015a; Toth, 2009; Zhang and Govindaraju, 2000). Despite their successes, there are still widespread criticisms on their black-box nature and cautions against their use in real-world problems in favor of physically sound conceptual models. To over- come this issue, recent efforts have been made by the research community to explain the internal workings of ANN, and link the processes taking place within the network to the processes in the watershed (Fernando and Shamseldin, 2009; Jain and Kumar, 2009; Jain et al., 2004; Wilby et al., 2003). Others have focused http://dx.doi.org/10.1016/j.jhydrol.2015.08.008 0022-1694/Ó 2015 Elsevier B.V. All rights reserved. ⇑ Corresponding author. E-mail address: cekwchau@polyu.edu.hk (K.-W. Chau). Journal of Hydrology 529 (2015) 1788–1797 Contents lists available at ScienceDirect Journal of Hydrology journal homepage: www.elsevier.com/locate/jhydrol